git branch --contains is slow with a lot of branches

Hi,

Sometimes I want to know what (possibly remote) branch contains a given commit. The repository where I do that has thousands of branches:

$ git for-each-ref | wc -l 7657

And a lot of commits:

$ git rev-list --all | wc -l 538174

Using git branch --contains can be a very expensive thing:

$ time git branch --contains 0812b94 --all > /dev/null

  real  3m0.871s
  user  3m0.828s
  sys   0m0.084s

I'd argue this shouldn't take much more time than enumerating all revs:

$ time git rev-list --all | wc -l 538174

real 0m4.842s user 0m4.488s sys 0m1.332s

This can be reproduced to a certain degree with the git git repo:

$ git clone https://github.com/git/git $ cd git $ for i in $(seq 1 1000); do git branch branch$i master; done $ git gc # will pack the refs $ time git rev-list --all | wc -l 40886

real 0m0.505s user 0m0.464s sys 0m0.108s

$ time git branch --contains v2.0.0 > /dev/null

real 0m6.207s user 0m6.204s sys 0m0.004s

(especially in this case where all branches point to the same commit)

It's also essentially linear on the number of branches:

$ for i in $(seq 1001 7000); do git branch branch$i master; done $ git gc $ time git rev-list --all | wc -l 40886

real 0m0.493s user 0m0.484s sys 0m0.076s

$ time git branch --contains v2.0.0 > /dev/null

real 0m43.446s user 0m43.436s sys 0m0.040s

Mike

On Tue, Mar 31, 2015 at 10:45:11AM +0900, Mike Hommey wrote:

Yes, this is well known. It does a separate traversal for each branch, which is why you noticed that it's linear in the number of branches.

I changed the "tag --contains" algorithm a while ago to do it all in one traversal. The downside is that it uses a depth-first approach which means it almost always goes to the roots. This is more appropriate for tags (as you often have old tags), but less so for branches.

I did some work on a contains() implementation that would is breadth-first, but handles multiple tips in a single traversal. It needs a little polish, and then to be hooked into "git branch". This is part of the proposed GSoC project for unifying "tag -l", "branch -l", and "for-each-ref".

-Peff